verification-before-completion

name: verification-before-completiondescription: Use when about to claim work is complete, fixed, or passing, or before committing or creating PRsmetadata:   openrig:     vendored_from: "Obra Superpowers (https://github.com/obra/superpowers)"    vendoring_pattern: modify-the-file    last_upstream_check: "2026-10-25 (SKILL.md and license compared at 2ce8aa15d05bd57213e6f746f953aafb7ded7a8b; no content sync)"    divergence_notes: |      Retained local copy; SKILL.md differs from the checked upstream revision.      This comparison did not refresh the skill body or companion files.      Upstream license terms are included in LICENSE.txt.

Verification Before Completion

Overview

Claiming work is complete without verification is dishonesty, not efficiency.

Core principle: Evidence before claims.

This rule serves shipping — it is not a ritual that outranks it. It governs one thing: claims that a verifiable result holds ("tests pass", "build succeeds", "bug fixed") — have evidence for the result before you assert it, and scale the check to the stakes (a one-line copy tweak is not a test-suite ceremony). Don't dodge it with synonyms; equally, don't inflate it into a tax on the word "done" or a reason to keep re-proving instead of shipping the working thing.

The Rule

Before claiming a verifiable result, inspect evidence that supports that result for the candidate and environment you name. Use an existing result while its relevant inputs and assumptions still hold; rerun the affected check when they change or are uncertain. State whether you ran the check or inspected another owner's evidence. A source read, syntax check or partial test supports only its stated scope. Ordinary progress reports, delegation and moving to another task do not require a new test run. Required project checks still apply.

A completed run is not automatically current, and unchanged source does not establish unchanged live state.

For a claim about performance, also read references/performance-claims.md.

Common Failures

Red Flags - STOP

• Using "should", "probably", "seems to"
• Claiming a specific result is proven ("tests pass", "it's fixed", "build's green") without evidence that applies to it — enthusiasm itself is fine; an unverified result claim is the flag
• About to commit/push/PR without verification
• Trusting agent success reports
• Claiming more than a partial check covers
• Thinking "just this once"
• Tired and wanting work over

Key Patterns

Tests:

✅ [Run test command, or read the named candidate's CI output] [See: 34/34 pass] "All tests pass"
❌ "Should pass now" / "Looks correct"

Regression tests (TDD Red-Green):

✅ Write → Run (pass) → Revert fix → Run (MUST FAIL) → Restore → Run (pass)
❌ "I've written a regression test" (without red-green verification)

Build:

✅ [Run build] [See: exit 0] "Build passes"
❌ "Linter passed" (linter doesn't check compilation)

Requirements:

✅ Check each promised outcome against applicable evidence → Report gaps or completion
❌ "Tests pass, phase complete"

Agent delegation:

✅ Agent reports success → Inspect the artifact and the evidence it names → Report what was checked and what wasn't
❌ Pass on the report alone

When the check itself lies — content over shape

A passing check is necessary but not sufficient: a check can exit clean and still prove nothing. These are the named ways a green check lies — the failure siblings behind "evidence before claims." (Deeper treatment of what you verify against: reference-first-verification.)

• Assert on CONTENT, never shape or count. Output that looks like success — a non-zero count, a clean exit, no error text — can carry failure as its content. Before trusting a check, ask what its output looks like when it fails: if failure and success are indistinguishable at a glance, the check proves nothing. ([1 lines] and No matches found both read as "a match" to a line-count.)
• Compare the returned value to the value you PASSED — never to a description of it. A confident label (evidenceRef NOW SET: printed above the untouched original) is indistinguishable from a verification in a transcript. Diff the actual result against your actual input, not against a claim about it.
• Find the untested axis. Thorough coverage of ONE axis reads as thorough coverage overall. Ask what every one of your tests shares — a healthy manifest, a default option value, the happy path — and treat that shared assumption as the dimension you never tested.
• A syntax check proves parseability and nothing else. A clean node --check / lint / parse pass is the covered axis; boot and runtime are the untested one. Parseable ≠ sound.
• A failure whose shape is a HANG reports nothing. A bare await against a stated deadline that never re-checks cannot fail loudly. Race it against a timer so failure has a shape you can see.
• A completeness inventory read through a truncating display is worthless. When deciding which occurrences to change, count with grep -c / grep -o / full-line output — never cut, head, or a pager, which can hide a match past the cut. Distinguish "I found N" (a count) from "I saw N" (a display). Audit multiline prose by paragraph, not by line — a line filter both misses long-line matches and false-flags paragraphs whose companion text is on the next line. Then re-audit with a different method than the one that built the inventory — same-method re-checking reproduces the blind spot. (Empty/short output is a claim about your COMMAND until re-proven a second way.)